Performance Evaluation

In this section, we provide the performance statistics with different combinations of delegate and buffer conversion methods for : Object Detection which is introduced above.

Designed Pipeline for Object Detection

Note

The statistics are roughly recorded on experimental. To obtain exact statistics, you should run the APP on the platform. Performance may vary between different versions of the board image.

Genio 350

The measurement data below was tested in performance mode. For information about how to set the performance mode on Genio-350, please refer to Genio-350 Performance Mode.

• Apply different conversion methods for the pipeline. This table only shows the statistics for applying NNAPI delegate, because it has the best performance of all based on the experimental results.

	USB Camera (resolution 1920*1080)		YUV camera (resolution 1920*1080)
Conversion Method	FPS	Inference Time(ms)	FPS	Inference Time(ms)
v4l2convert + v4l2convert	14	34	19	30
videoconvert + v4l2convert	14	34	4	33
v4l2convert + videoscale	5	34	5	31
videoconvert + videoscale	16	32	3	31

• Apply different delegates for the pipeline with the best conversion methods. For USB Camera, is videoconvert + videoscale, and for YUV camera, is v4l2convert + v4l2convert.

Average inference time(ms)

Object Detection	USB camera (640*480)		YUV camera (640*480)
Delegate	FPS	Inference Time(ms)	FPS	Inference Time(ms)
CPU	4	254	4	255
GPU	6	174	5	172
ArmNN(GpcAcc)	12	82	9	75
ArmNN(CpcAcc)	11	90	9	75
NNAPI(VPU)	24	34	14	33
Image Classification	USB camera (640*480)		YUV camera (640*480)
Delegate	FPS	Inference Time(ms)	FPS	Inference Time(ms)
CPU	8	137	7	131
GPU	12	84	10	82
ArmNN(GpcAcc)	21	43	15	42
ArmNN(CpcAcc)	18	50	16	41
NNAPI(VPU)	31	19	23	18

Genio 1200

The measurement data below was tested in performance mode. For information about how to set the performance mode on Genio-1200, please refer to Genio-1200 Performance Mode.

• Apply different conversion methods for the pipeline. This table only shows the statistics for applying ARMNN(CpuAcc), because it has the best performance of all based on the experimental results.

	USB Camera (640*480)
Convert Method	FPS	Inference Time(ms)
v4l2convert + v4l2convert	Unsupported yet
videoconvert + v4l2convert
v4l2convert + videoscale
videoconvert + videoscale	30	30

• Apply different delegates for the pipeline with the best conversion methods. For USB Camera, is videoconvert + videoscale.

Average inference time(ms)

Object Detection	USB camera (640*480)
Delegate	FPS	Inference Time(ms)
CPU	31	29
GPU	26	33
ArmNN(GpcAcc)	18	18
ArmNN(CpcAcc)	30	30
Neuron(MDLA)	31	8
Image Classification	USB camera (640*480)
Delegate	FPS	Inference Time(ms)
CPU	31	16
GPU	31	15
ArmNN(GpcAcc)	31	12
ArmNN(CpcAcc)	31	17
Neuron(MDLA)	31	3

Note

• For how to convert tflite model to dla model, please refer to Neuron Compiler section.

• YUV camera: Unsupported yet

Genio 700

The measurement data below was tested in performance mode. For information about how to set the performance mode on Genio-700, please refer to Genio-700 Performance Mode.

• Apply different conversion methods for the pipeline. This table only shows the statistics for applying ARMNN(CpuAcc), because it has the best performance of all based on the experimental results.

	USB Camera (640*480)
Convert Method	FPS	Inference Time(ms)
v4l2convert + v4l2convert	Unsupported yet
videoconvert + v4l2convert
v4l2convert + videoscale
videoconvert + videoscale	31	20

• Apply different delegates for the pipeline with the best conversion methods. For USB Camera, is videoconvert + videoscale.

Average inference time(ms)

Object Detection	USB camera (640*480)
Delegate	FPS	Inference Time(ms)
CPU	30	31
GPU	19	47
ArmNN(GpcAcc)	28	29
ArmNN(CpcAcc)	31	20
Neuron(MDLA)	31	9
Image Classification	USB camera (640*480)
Delegate	FPS	Inference Time(ms)
CPU	31	17
GPU	31	23
ArmNN(GpcAcc)	31	13
ArmNN(CpcAcc)	31	13
Neuron(MDLA)	31	4

Note

• For how to convert tflite model to dla model, please refer to Neuron Compiler section.

• YUV camera: Unsupported yet